Home / Tools / Which Nova do I need?
Tool 01 · Model sizer

Which Nova
do I need?

Cooled by Nexalus

Pick a model and how you'll use it. We'll work out the GPU memory it needs and the Nova server that fits it.

Memory maths uses each model's published architecture. Throughput is scaled from our one measured result (Llama 3.3 70B, FP4, up to 55,000 tok/s per server); everything else is an estimate.

Recommended

—

Explore

Memory needed vs one server

Nova 8 Inference · 768 GB
Nova 8 Training · 1,128 GB
Weights
KV cache
Optimiser
Total

How the 8 GPUs are used

More tools

Keep exploring.

Cooled by Nexalus